Driving Style Classification for Vehicle-Following with Unlabeled Naturalistic Driving Data

Abstract

Driving style has great influence on fuel economy and safety and plays an important role in intelligent transportation and vehicles. A driving style classification methodology with unlabeled naturalistic data is proposed to shed light on drivers' vehicle-following characteristics, and the driving style model is presented to classify and keep the driver's own style and to improve fuel economy and safety. Firstly, NGSIM dataset of NHTSA is chosen as unlabeled data and its Gipps model parameters are calibrated via genetic algorithm to represent vehicle-following characteristics. Then, the Gipps model parameters sensitivity are studied and the expected maximum ego vehicle deceleration and the expected maximum preceding vehicle deceleration are set as the drive style index, and the vehicle-following driving style is classified into risky, normal and conservative type depending on the Silhouette coefficient via the DBSCAN cluster. Finally, the vehicle-following driving style model is proposed by the cluster centroids and boundaries. Simulation of WLTC and cut-in scenarios manifest that the proposed driving style classification methodology can classify drivers' style online and can be used for human-centered driving control, improving vehicle-following safety, traffic efficiency, fuel economy and humanity.